Relation of genomic variants for Alzheimer disease dementia to common neuropathologies

Participants.

Data were obtained from the Religious Orders Study (ROS) and the Rush Memory and Aging Project (MAP), both community-based clinical-pathologic cohort studies of aging and dementia. Participants were free of known dementia at baseline and agreed to annual clinical evaluations and brain donation. ROS enrolls older Catholic priests, nuns, and brothers from across the United States since 1994. MAP enrolls older residents from northeastern Illinois since 1997. The follow-up participation rate exceeds 95% and the autopsy rate exceeds 85%. Detailed information on ROS and MAP design have been published previously.^26,27

At the time of these analyses, 1,412 of 3,063 participants enrolled had died, and 1,221 had undergone autopsy (autopsy rate = 86.5%). Of these, genome-wide genotype data were available on 1,079 participants. We excluded participants with neuropathologic data that had not yet been completed, resulting in a total of 1,017 participants. The average age at death was 88.6 years with SD of 6.5 years; 649 (63.8%) were female; and the average years of education was 16.4 (SD 3.6).

Standard protocol approvals, registrations, and patient consents.

The studies were approved by the Institutional Review Board of Rush University Medical Center. All participants gave written informed consent and signed an Anatomic Gift Act for organ donation.

Postmortem procedures.

Removal and processing of the brain followed a standard protocol used in all locations. The mean postmortem interval was 8.3 hours (SD 7.4 hours). Neuropathologic evaluations were performed at the Rush Alzheimer's Disease Center, blinded to clinical data, and reviewed by a board-certified neuropathologist.²⁸

Neuritic plaques (NP) and neurofibrillary tangles (NFT) were identified with modified Bielschowsky stain in 6-μm sections in 5 different regions (entorhinal cortex, hippocampus at CA1, midfrontal cortex, middle temporal gyrus, and inferior parietal cortex). Each case received a semiquantitative estimate that assessed the burden of NP as recommended by Consortium to Establish a Registry for Alzheimer's Disease and a Braak score that assessed the distribution and severity of NFT pathology.^29,30 The final postmortem diagnosis of pathologic AD was based on the recommendations of the National Institute on Aging (NIA)–Reagan criteria.³¹ The diagnosis required an intermediate or high likelihood pathologic AD. Lewy bodies (LB) were identified by α-synuclein immunohistochemistry in one or more neocortical regions (midfrontal, middle temporal, and inferior parietal) together with either midbrain or limbic LB.³² Macroinfarcts were recorded from gross examination of fixed slabs and were confirmed by histology. Microinfarcts were documented with hematoxylin & eosin.³³ Hippocampal sclerosis was assessed in a hemisection of the mid-hippocampus and defined as neuronal loss and gliosis in CA1/subiculum or other sectors.³⁴

Genotyping.

DNA from ROS and MAP participants was extracted from peripheral blood mononuclear cells or frozen postmortem brain tissue. Genotyping was done in 2 phases. The data from the first phase (n = 889) were generated on the Affymetrix Genechip 6.0 platform at either the Broad Institute's Center for Genotyping or the Translational Genomics Research Institute. The data from the second phase (n = 128) were generated on the Illumina HumanOmniExpress at the Children's Hospital of Philadelphia. These 2 sets of data underwent the same quality control (QC) analysis using PLINK software.³⁵ We applied standard QC measures for participants (genotype success rate >95%, genotype-derived sex concordant with reported sex, excess interheterozygosity/intraheterozygosity) and for single nucleotide polymorphisms (SNPs) (Hardy-Weinberg equilibrium p > 0.001; minor allele frequency >0.05, genotype call rate >0.95; mishap test >1 × 10⁻⁹). Subsequently, EIGENSTRAT³⁶ was used to identify and remove population outliers using default parameters. Imputation was performed using the BEAGLE software (version 3.3.2). We used reference haplotype panels of Northern European ancestry (CEU) from the 1000 Genomes Project (The 1000 Genomes Project Consortium Interim Phase I Haplotypes, 2010–2011 data freeze).

The susceptibility variants interrogated in this study have been reported to have genome-wide significant associations with AD dementia.^17–23 The variants in CR1, BIN1, CD33, CLU, ABCA7, CD2AP, PICALM, EPHA1, MS4A6A, and MS4A4A were reported in earlier GWAS. The variants in HLA-DRB5, PTK2B, SORL1, SLC24A4, DSG2, INPP5D, MEF2C, NME8, ZCWPW1, CELF1, FERMT2, and CASS4 were identified in the most recent large-scale meta-analysis with a sample size of nearly 75,000.²¹ Detailed information of variants is included in table e-1 on the Neurology® Web site at Neurology.org. We replaced rs3764650 (ABCA7) with rs3752246 because the original index SNP had poor imputation quality in our data. APOE was directly genotyped by sequencing codon 112 and codon 158 of exon 4 of the gene.³⁷

Statistical analysis.

We used multiple logistic regression analyses to examine the association of genomic variants with pathologic AD. In these analyses, the pathologic AD diagnosis of intermediate- or high-likelihood AD by NIA Reagan criteria was used as the binary outcome and we modeled the odds of having pathologic AD regardless of the clinical diagnosis. First, we included APOE as a positive control, where we confirmed that the presence of any ε4 allele was associated with greater odds of having pathologic AD and the presence of any ε2 allele was associated with lower odds of pathologic AD, as we have reported previously.³⁸ Next, we repeated the analyses to examine each of the other 22 genomic variants separately. Here, the dosage for each variant was used as the main predictor. The dosage was imputed assuming additive inheritance and was in the form of a continuous variable ranging from 0 to 2. Therefore we chose additive models such that the corresponding regression coefficient, after taken exponential, estimates the odds ratio (OR) of pathologic AD for each additional copy of the coded allele. All the models were adjusted for age at death, sex, and the first 3 principal components from EIGENSTRAT in order to minimize potential confounding due to subpopulation stratification. Similar regression analyses were performed to examine the associations with the presence of other common neuropathologies. A Bonferroni-corrected α level of 0.002 was used for statistical significance throughout the analysis. This threshold was calculated as 0.05 divided by 24, the total number of susceptibility variants. The association analyses were performed using SAS software version 9.3 (SAS Institute, Cary, NC).

We next evaluated the statistical power by estimating the total sample size required to reach genome-wide significance (p < 5 × 10⁻⁸) with 80% power in a 2-tailed test. The sample size calculation depends on the case-to-control ratio, frequency of the variant (with respect to the coded allele), baseline disease risk, as well as the estimated effect size. We obtained these parameters based on our data as well as the results from the association analyses. Specifically, the case-to-control ratio was estimated by the ratio of total numbers of cases with pathologic AD diagnosis and those without. The frequency of the coded allele was used to determine the proportions of copies for each variant. The baseline risk of pathologic AD was the estimated probability of AD for a typical woman with mean age 89 years and absence of the coded allele. The effect size was the estimated OR with respect to the coded allele. These post hoc power analyses were performed using the Quanto program (version 1.2.4). We examined the variance of individual variants explained by the model covariates (R²), which may also influence the sample size estimation.³⁹ In our data, these R² were minimal and ranged from 0 to 0.017 with a median of 0.002. Therefore, the estimated sample size will be slightly larger if correcting for the R²s but not at a level sufficient to alter our conclusions.

Associations of APOE with pathologic AD.

We first examined the association of APOE with pathologic AD. A total of 25.6% of the autopsied cases had the ε4 haplotype. The presence of ε4 haplotype almost quadrupled the odds of having pathologic AD and easily met genome-wide significance (OR 3.82, 95% confidence interval [CI] 2.67–5.46, p = 1.9 × 10⁻¹³). Additionally, 15.5% of cases had the ε2 haplotype. The presence of ε2 allele was associated with more than a 50% lower odds of having pathologic AD, although the association did not meet genome-wide significance (OR 0.42, 95% CI 0.30–0.608, p = 3.1 × 10⁻⁶). To illustrate the additional power gained with pathologic AD, we compared the results of the associations with pathologic AD to that of AD dementia. Here, the presence of ε4 haplotype triples the odds of having AD dementia (OR 3.09, 95% CI 2.26–4.25, p = 2.5 × 10⁻¹²), and the presence of ε2 haplotype was associated with about a 40% lower odds of having AD dementia (OR 0.63, 95% CI 0.43–0.92, p = 0.017). In both cases, the strength of the association and the p value are markedly attenuated relative to pathologic AD, as previously reported.³⁸ This supports our hypothesis that genomic variants for AD dementia that are also genomic variants for pathologic AD should be found with smaller samples sizes and that the absence of such a finding suggests that the variant may be associated with another pathology or other factors associated with AD dementia.

Associations of other genomic variants with pathologic AD.

Next, we examined the association of 22 additional AD dementia genomic variants with the odds of having pathologic AD (table 2). We found 1 variant in the CR1 gene where the G allele of rs6656401 was associated with a lower odds of pathologic AD (OR 0.62, p = 0.0003), as previously reported.^24,25 In addition, the associations of 3 other variants with pathologic AD reached p < 0.05. Specifically, the T allele of rs1476679 in the ZCWPW1 gene was associated with greater odds of pathologic AD (OR 1.37, p = 0.002), as previously reported.²⁴ The minor allele (T) from rs35349669 of the INPP5D gene was associated with a greater odds of pathologic AD (OR 1.31, p = 0.009) and the major allele (T) of rs17125944 in the FERMT2 gene was associated with a lower odds of pathologic AD (OR 0.71, p = 0.050). The directions of these associations are consistent with the findings from the recent GWAS analysis on AD dementia. By contrast, the remaining 18 previously reported genomic variants for AD dementia had ps > 0.05 with our sample size.

Associations of genomic variants with other neuropathologies.

We repeated the analysis to examine potential associations of these genomic variants with presence of cerebrovascular infarcts, LB, and hippocampal sclerosis (table e-1). The results are illustrated in figure 2. While none of these associations survived stringent Bonferroni correction, we found some evidence that a few variants may be associated with non-AD neuropathologies. Briefly, the C allele of rs3865444 in the CD33 gene was associated with lower odds of having macroscopic infarcts (OR 0.809, p = 0.037) and the T allele of rs10838725 in the CELF1 gene was associated with greater odds of having microinfarcts (OR 1.371, p = 0.005). Interestingly, the same variant in CELF1 was associated with lower odds of LB (OR 0.764, p = 0.019). The presence of APOE ε4 was associated with greater odds for having hippocampal sclerosis (OR 2.396, p = 0.0006), as was the T allele of rs11767557 in the EPHA1 gene (OR 1.728, p = 0.029).

Sample size estimation to reach genome-wide associations.

We next estimated the sample sizes needed for all of the genomic variants to achieve genome-wide significance with pathologic AD (figure 3, truncated to n = 75,000). Given the model outlined in figure 1, we expect that variants associated with AD dementia that are also associated with pathologic AD should be identifiable with a sample size less than 75,000, the size of the most recent AD dementia GWAS. We illustrate our calculation by using the CR1 variant (rs6656401). Based on the percentage of pathologic AD in our sample (62.9%), we assume that for every 100 AD cases there are 60 controls. The coded allele (G) had an estimated allele frequency of 0.81. According to the logistic regression result, the estimated probability of having pathologic AD for a typical woman who died at age 89 and was rs6656401^A/A homozygote was 0.81. Our power calculation shows that, at α level of 5 × 10⁻⁸ and 80% power, we need a minimum of 2,614 participants to detect an additive effect as large as 0.62, the OR obtained from our regression model.

We repeated this post hoc power analysis for the remaining variants (table e-2). The results are illustrated in figure 3. As expected, the only variant detectable with our current sample size is the APOE ε4. Four variants, including APOE ε2, CR1, INPP5D, and ZCWPW1, could be detected with a sample size of 5,000 or less. Overall, besides APOE, 11 of the 22 interrogated variants were detectable with sample sizes up to 75,000. The results suggest that half of the previously reported genomic variants associated with AD dementia are not likely to be genomic variants for pathologic AD. Figure e-1 illustrates how many genomic variants are detectable with sample sizes 2×, 3×, 4× or more the size of the largest GWAS for AD dementia (truncated at n = 1,000,000).